protologue

Constitutional Classifiers

Constitutional classifiers are input and output safeguard classifiers trained on synthetic data generated from a written constitution that defines permitted and restricted content, designed to block universal jailbreaks.

Description

Because the training data is generated from natural-language rules, the classifiers can be updated as threats change by editing the constitution. Their authors reported that thousands of hours of red teaming found no universal jailbreak against the protected system, at a modest cost in refusals and compute.

Sources

  1. Sharma et al. (2025). Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming.

Cite this entry

Protologue. (2026). Constitutional Classifiers. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0168). https://protologue.com/t/constitutional-classifiers/

BibTeX
@misc{protologue_constitutional_classifiers,
  title = {Constitutional Classifiers},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0168},
  url = {https://protologue.com/t/constitutional-classifiers/}
}

Markdown JSON